I’m a core developer for the Tesla OpenPilot Fork and have driven over 30,000 miles with OpenPilot. It works better than AP1 for everything but stop & go traffic, and it only fails there because pre-AP tesla’s dont have an iBooster to control the friction brakes, so our fork has to use the regenerative brakes which only slow the car down to ~6 mph.
The current model is great, steering controls (on a pre-ap tesla) is much smoother than AP1 or AP2+. If you look at a Tesla on Autopilot’s steering wheel when doing long and sharp turns, you will see it jerk quite a lot, OP has none of that in their current model. The only issue it really has is some cars don’t have EPAS’ with enough torque for some sharper turns. The new 2019 corolla, that is officially supported, and obviously Tesla which have the same EPAS as new teslas, have a strong enough EPAS to make every turn.
While the current model is extremely good, it’s based on current machine learning practices of label everything, build models, test what is produced. Not much different than every other self driving company out there, but definitely the best executed experience thus far when it comes to highway driving. hands-free perfect steering with driver monitoring just cannot be beaten.
However… Comma has some seriously advanced ML Stuff happening internally which they want to release in the future (maybe v1.0.0?) that is similar to Tesla’s “Dojo”, and in that aspect are years ahead of most self driving companies. When he talks about “laneless” models, It’s pure unsupervised learning with no labeling of lines or cars or anything, Similar to how alphago learns by playing against itself, the model learns from driving against 12 million miles of real drivers.
Knowing the George and the Comma team, some of who are far better than George at ML, I’m extremely confident that they will be able to pull it off too, even if they are a couple years behind Tesla in doing so.